Hello! Welcome to my portfolio. I hope you enjoy my findings of this class corpus.
A friend and I made a few songs in the Christmas Holidays in 2024. I wanted to use these songs, because of how much time we spend making them in my friends basement, where we never saw the light of day.
I actually wrote the first song ‘lennart-p-1’ in high school for an assignment. The real title of this song is “Feeling Free”, which is based on the feeling of a late night bicycle ride. We recorded this in her basement. She plays the keys and I play guitar, which is why these will be the most dominant instruments. The song has a long build-up using reoccurring themes which eventually harmonize together. The ending is somewhat faster and more aggressive, to create some contrast in the song, to add some spice!
We wrote the second song ‘lennart-p-2’ together in the basement. This songs real title is “Retake The Long Day”, which is based on the many retakes and long days of recording. It started somewhat as a pop song, but gradually changed to some kind of rock symphony? I like to look at it that way. We wanted to express our skill in the sounds of the instruments and technique in the solo’s. We put many hours in this one.
I recommend to listen to the songs while looking at the grams and descriptions.
Enjoy!
In this graph you can see the approachability and engagingness. I thought there could be a correlation. I hypothesized that the higher the approachability, the higher the engagingness should be. This was wrong, there is no clear sign of a correlation in this graph. So I think I don’t really know what is meant by approachability. I do, however, see a correlation between engagingness and arousal. It looks like the higher the engagingness, the higher the arousal is. Which makes sense when it comes to mind. lennart-p-1 and lennart-p-2 are pretty close to each other.
These are the chroma features of the class corpus. I compared both my tracks’ features individually, to see the difference. Both the songs are not made by AI. The most noticeable difference is in tempo, but this was actually a mistake by the features. The tempo of the second song is 105 BPM, not 152. The rest of the features are actually quite close.
This spectogram shows the chromagram of ‘lennart-p-1’. From this chromagram we can see that most of the song was in D, because D, the tonic, F#, the third and A, the fifth. The C# can be explained because the chord Dmaj7 is played a lot. The B shows up pretty often as well. This is probably because of the frequent occurrence of E7.
Here we can see the timbre of the first song. Noticeable is the beginning and the last third of the song. The beginning starts only with guitar, which could be the reason for that difference in color. The last third really different from the rest of the song. It’s more aggressive in sounds (more distortion) and faster in tempo. This switch was made with a break, which can be seen after 150, where it is very yellow.
This song is in G minor, because of the G, Bb, D, or in Eb major, because of Eb, G and Bb. It is a bit tricky, but I do think the key is G minor. The ending is a long open ending on G and D, which is a G power chord. I must say, why C, Eb and A occur this much is unclear to me, even though I was one of the two writers.
There is a silence after the intro. This silence is portrayed a clearer yellow just before 100 seconds.
Here you can see some structure of the song. There are a guitar solo’s both at the more yellow areas around; One around 50 seconds and one around 160 seconds. Also the ending is quite noticeable.
Here you can see the structure of the second song. Around 50 seconds, you can see the first transition to another song segment. Just after 100 seconds, you can see the next transitioning. From here on, the song changes gradually, as I mentioned in the introduction. Around 240-250 seconds you can see a somewhat brighter color. At this point in time, the guitar and Hammond solo’s take place.
In this matrix you can see the intro and the more aggressive faster longer outro. The rest seems somewhat the same.
In this matrix you can see the intro at 0 till 20 seconds, the next segment around 60 seconds, and from 100 the same timbre. I think that overall the chroma-based self-similarity matrices showed the structure of the songs better in this case. So the songs structures are more based on the difference in pitches than in timbre. ***
This song is Hey Joe from Jimi Hendrix, a classic!
Both the matrices have a similar structure. This is because the song has an overall simple structure. It’s the same chord scheme over and over again. The only things that change is the switches between singing and playing the guitar. But that doesn’t seem enough to really stand out on the matrices.
The key of this song is mostly D major and F# minor. They switch very often. I sometimes used chords like D7, which is outside the original key. The gram shows that also A major is a frequently used key. I have to disagree, because when I play A, I play A7.
The second songs keygram is much clearer. The key is G minor.
This is the chordogram of the first song. Even though it seems very random, it is a little bit accurate. The chords till around 60 seconds are Dmaj7, E7, A7, F#min7, D7 and C#min7. The second part from 60 seconds to 85 contains F#min7, E7, Dmaj7, C#min7, C#7, E7 and Bmin7.
The only part this chordogram is only a bit certain in the bridge around 55 seconds to 100. Which is strange, because this song has simpler chords than the previous one. I think it is so uncertain because of all the motives that are played op on it.
In this histogram you can see that there is a high peak at 150 BPM, which is really fast. I think this is the result of mistakes of tempo determination.
The Energy Novelty shows the difference in energy compared to the previous moment in the song. The peak around 20 seconds can be explained by the addition of the bass, drum and lead guitar. The peak around 155 can be explained by the beginning of the second guitar solo. Just before the solo is a silence, which can be seen in the gram. Than the solo begins, which is a intense difference from the previous silence. The smaller peak in the middle around 90 seconds is the beginning of the theme after the first guitar solo. This solo ends with a long ending note, than comes both the themes simultaniously.
In the second song are more peaks than the previous song, but these peaks are less obvious. The first peak is just the start of the song. The second peak is the drum and rhythm guitar that joins. There is a kind of silence at 58 seconds and a new part of the song after that. The peak just before 200 seconds is randomly towards the end of the guitar solo. I can’t explain this one. Maybe one screeching bend screaming above all the other parts…?
The Spectral Novelty compares the one window of music with the previous window and shows if the difference is high or low. The first song shows little structure. The only really noticeable things are the break just after 150 seconds and the ending after 250 seconds. This last one is actually interesting, because it seems to decline. This can be explained by the drum leaving early, and being very repetitive afterwards, except for the final blow!
The structure of the second quite more interesting. At 20 seconds, the drum, bass and rhythm guitar start playing. At 58 seconds, the intro stops and at 62 seconds the next more calm part starts. At around 98 seconds, the violins (played on keyboard) that join, can be seen. Also, at the end you can see a decline, because the ending is kind of repetitive.
The tempo of the first song looks like to be 90 BPM. Which is very correct! The shift in the gram around 155 seconds is very noticeable. Here is a tempo change! The BPM of this second part of the song is 109 BPM.
The second song is 105 BPM. Here is no tempo change.
In this cyclic tempogram of the first song, you can see the shift as well.
In this cyclic tempogram of the second song, you can see that the tempo is more spiky. This is probably because the first song was actually not a real drum. It was a basic tempo played by the recording software. The second song was recorded on a real drum, played by a real drummer. Interesting to see that the difference was recognized!
lennart-p-1 stands between sanne-o-1 and cecilia-b-2, and lennart-p-2 between elze-s-1 and tobias-p-2. Both of them are far on the right. lennart-p-1 doesn’t sounds anything like sanne-o-1 and cecilia-b-2, but I guess they come the closest out of the whole class corpus? lennart-p-2 has some similarities with the other songs. elze-s-1 and tobias-p-2 are very rock, but a different kind of rock than lennart-p-2. Their songs are much heavier than my pop to rock symphony song. Their songs are real bangers though.
Truth
Prediction AI Non-AI
AI 35 21
Non-AI 14 20
# A tibble: 2 × 3
class precision recall
<fct> <dbl> <dbl>
1 AI 0.625 0.714
2 Non-AI 0.588 0.488
# A tibble: 2 × 3
class precision recall
<fct> <dbl> <dbl>
1 AI 0.646 0.633
2 Non-AI 0.571 0.585
The precision and recall are as high is you would like it to be. AI gets classified wrongly too many times to be somewhat reliable.
This shows the importance order of the features of this class corpus.
In this plot, you can see that the higher the ‘instrumentalness’ is, the lower the ‘engagingness’ is. Tempo and AI seems irrelevant to this correlation. The instrumentalness of my songs are quite low, for songs that are entirely instrumental.
I hope my findings were insightful and interesting. I learned many ways to analyze music on this low level (digital audio), instead of high level (e.g. sheet music analysis). I learned to analyze on all 4 ‘moments’ in music: pitch, volume, timbre and tempo. I liked how well the similarity-matrices worked to analyse the structure of the songs. Also the novelty functions seemed to point out interesting changes in the songs. Also very interesting that the chroma features of the AI songs didn’t seem to be too different from the non-AI songs. Shout out to my friend Joleine who helped writing the second song and shout out to fellow students who helped me with some issues. Thank you for reading!